The goals system

I wanted my AI to actually know me, not just remember the last few messages of a conversation. So I built what I call Telos, adapted from an idea in Daniel Miessler's PAI project: a structured model of who I am and what I am trying to do. Plenty of people write their goals down; the difference here is that the system also watches what I actually did with my month, and tells me about it.

What goes in.

• A vault sweep. The first draft came from pointing Claude at years of my own essays, journals, and notes and asking it to assemble who I am from what I had already written: sourced rather than aspirational.
• An interview. Claude asked, I talked, the transcript got distilled. It filled in everything I had never written down because nobody had asked.
• A continuous dictation log. Any self-reflective thought, straight from my phone, appending without approval, because friction on capture kills capture.
• A quarterly revealed-action pass. What I actually did rather than what I said: where the calendar time went, which tasks I finished and which I quietly killed. The route built to catch self-deception.
• The conversations themselves. Working a decision through the system surfaces patterns in how I decide that were never in any file, and those get proposed back.

The files. A collection of files detailing: me (my identity, my relationships, my beliefs and models, my strengths and weaknesses); my purpose (my mission, my goals, my strategies); and my learnings. Each goal feeds a mission and each strategy feeds a goal, so a strategy serving no goal is visible at a glance. Behind every summary file sits a raw log: near-verbatim, dated, never edited after landing. Logs append without my approval; nothing reaches a summary without it. Claude may suggest additions to my missions, goals, and the rest, but nothing lands without being run past me first.

What comes back out.

• Claude reads the summaries at the beginning of every session that touches my own work, which means it catches me at my known failure points: if I am refining a model because the modelling is fun, it tells me to stop refining and go test the thing in the real world.
• It reads my task system back as evidence: what I opened, what I finished, and what I quietly killed, tested against what I said I was trying to do.
• Every month it pushes me a review of how I have actually spent my time against my strategies, which feed my goals: an audit of what is quietly being swept under the rug. It arrives already written, findings and proposed edits included, so I only have to rule.
• Coaching on demand, for decisions big and small. It reads the relevant raw logs, not just the summaries, and I can combine it with the consultants I have built on specific corpuses: the LessWrong one makes the Sequences available as a personal life coach equipped with a deep psychoanalysis of my brain. Every session writes a record, and the review reads those first, hunting one pattern in particular: you have brought me this decision three times in two months and still not made it.

I wrote separately about what it means to wire an AI into the authoring of your own goals, and where I think the limits are: The Extended Self.